Multiple-point simulation algorithms based on texture synthesis strategies cannot be directly parallelized. These algorithms simulate the category of visited nodes in a grid by visiting them in an ordered path. At each node, the previously simulated points in a fixed template called ?causal template? are taken into account, as well as some forthcoming nodes within the so called ?non-causal template?, in order to impose conditioning of hard data.
Simulation cannot be parallelized by simply dividing the field and simulating regions independently. However, the fixed path used in the implementation makes possible several strategies of parallelization, since simulation proceeds row-wise, and at some point along each row, there are enough simulated points within that row to be used as conditioning of the next row in the causal template. Therefore, a new process can start simulating concurrently the nodes within that row.
Depending on the size of the realization and of the template, this can be generalized and several processes can work concurrently. We present the implementation details of this approach and show some examples run in a cluster, along with the speed ups reached. |
